To chain two flaws and take over the ChatGPT and Codex accounts of several OpenAI employees, three researchers at the security firm Hacktron utilized Anthropic’s Claude Opus 5. The chain began with a bug in OpenAI’s public help forum software and moved through a weakness in the login system. This security research led to internal access within 72 hours, which OpenAI confirmed and fixed promptly after the report. Hacktron received a bounty for their findings.
The researchers pointed out an identity problem within OpenAI’s login system, not the forum software, which allowed the bug to reach staff accounts. The flaw in the forum software, Discourse, was related to image processing vulnerabilities that the researchers exploited with the help of AI models like Claude Opus 5.
The researchers used AI models to automate the exploitation process, with Claude Opus 5 proving to be successful in creating a working exploit. This aligns with the trend of AI models reducing the time and expertise required for offensive activities. The broader project, HEIF Heist, uncovered similar image-decoding flaws in software used by other companies, showcasing the impact of such vulnerabilities.
To mitigate risks associated with image-processing vulnerabilities, it is recommended to update libheif to the latest security release, limit trust in single sign-on services, and implement additional security measures for sensitive actions. Despite the findings, there is no evidence of real-world exploitation of the OpenAI flaw.
Overall, this incident highlights the importance of addressing vulnerabilities promptly and implementing robust security measures to safeguard against potential exploits.
